Advancing hydrogen storage: Explainable machine learning models for predicting hydrogen uptake in metal-organic frameworks

Abstract

· Explainable machine learning models were developed to predict hydrogen storage capacity in MOFs using a dataset of 1,729 experimental records. · The Extra Trees algorithm demonstrated superior predictive accuracy, achieving an R² of 0.995, RMSE of 0.1445, and MAE of 0.0762. · Outlier detection using the leverage approach confirmed that approximately 98% of the dataset falls within the model’s applicability domain. · The physical validity and interpretability of the developed model was testified using trend analysis and SHAP method. · The Bayesian regularized neural network (BRANN) model's explicit formulation clearly elucidates the mathematical relationships between inputs and outputs. · The explicit nature of the AI-based BRANN correlation enhances its utility and practicality for engineering applications. Metal organic frameworks (MOFs) exhibit exceptional efficacy in hydrogen storage owing to their distinctive characteristics, including elevated gravimetric densities, rapid kinetics, and reversibility. An in-depth look at existing literature indicates that while there are many studies using machine learning (ML) algorithms to develop predictive models for estimating hydrogen uptake by MOFs, a great number of these models are not explainable. The novelty of this work lies in the integration of explainability approaches and ML models, providing both accuracy and interpretability, which is rarely addressed in existing studies. To fill this gap, this paper attempts to develop explainable ML models for forecasting the hydrogen storage capacity of MOFs using three ML techniques, including Bayesian regularized neural networks (BRANN), least squares support vector machines (LSSVM), and the extra tree algorithm (ET). An MOF databank comprising 1729 data points was assembled from literature. Surface area, temperature, pore volume, and pressure were employed as input variables in this database. The findings demonstrate that of the three algorithms, the ET intelligent model attained exceptional performance, yielding precise estimates with a root mean square error (RMSE) of 0.1445, mean absolute error (MAE) of 0.0762, and a correlation coefficient (R 2 ) of 0.995. In addition, a novel contribution of this study is the generation of an explicit formula derived from BRANN, enabling straightforward implementation of hydrogen storage predictions without requiring retraining of complex models. The sensitivity analysis employing Shapley Additive Explanation technique revealed that pressure and surface area were the most significant features influencing hydrogen storage, with relevance values of 0.84 and 0.59, respectively. Furthermore, the outlier detection evaluation using the leverage method showed that approximately 98% of the utilized MOFs data are trustworthy and fell within the acceptable range. Altogether, this work establishes a distinctive framework that combines accuracy, interpretability, and practical usability, advancing the state of predictive modelling for hydrogen storage in MOFs.

Topics & Concepts